This paper tackles intra-class variations in generalized few-shot semantic segmentation via prototype learning, where appearance discrepancies between support and query targets degrade prototype effectiveness and new-class generalization. We propose the Relevant Intrinsic Feature Enhancement (RIFE) method, which leverages: (1) a feature decoupling and recombination mechanism to enhance support-query semantic consistency by mining robust features against intra-class variance; (2) a joint learning strategy that explicitly segments both support and query images during inference to verify prototype discriminability and refine segmentation results. Extensive experiments on \(\mathrm {PASCAL-5}^i\) and \(\mathrm {COCO-20}^i\) demonstrate that our approach surpasses state-of-the-art methods.